連接網絡統合分析中斷裂證據網絡之統計模型

Statistical Approaches to Joining Disconnected Evidence in Network Meta-analysis

宋成潔(Chengjie-Song)

指導教授 ： 杜裕康

國立臺灣大學/公共衛生學院/流行病學與預防醫學研究所/碩士(2017年)

https://doi.org/10.6342/NTU201702821

摘要

研究背景
相較於傳統統合分析，網絡統合分析不僅能同時比較多種處理，還能間接估計從未被直接比較過的兩種治療效果，各種比較相互連接形成網絡，更能滿足現代實證醫學的需求。
在統計方法不斷完善的過程中，研究者都希望能最大化地利用臨床試驗的數據，得到更可信的結論。可是在臨床試驗的設計中，常會因為一些現實因素導致網絡統合分析的網絡“斷裂”，這些斷裂的資訊便無法被經典網絡統合分析納入到證據體系中加以利用。導致網絡統合分析“斷裂”的情況如：纳入因研究成本、臨床試驗危險性以及受試者數目限制等因素而產生的single arm trial；因科學技術的突破性進展而產生了“典範轉移”，新藥之間彼此比較，舊藥之間彼此比較，中間缺少新藥與舊藥相互比較的連接機制；因不同地區醫療水平限制而產生不同的標準治療；某些研究者報告了不同於其他研究的評估指標、持有未發表的數據…以上因素都會造成證據網絡的斷裂。
針對網絡“斷裂”的情況，研究者可以試圖聯絡作者以取得未發表的數據，或者等待新的臨床試驗結果公佈。在以上途徑皆無法連接“斷裂”的網絡時，研究者即可訴諸不同模型來利用斷裂的證據。

研究目標
本研究的目的在於比較不同網絡統合分析的模型，合理地將斷裂的網絡重新連接起來，并結合模擬數據與實際資料比較連接結果，得到不同模型與方法的適用條件。

研究方法
本研究選擇三種網絡統合分析模型進行比較，分別為Lu Ades mode，Baseline and treatment effect model與Arm-based model。這三種模型背後的隨機效應假設有一定的差異：Lu Ades model為現今經典的網絡統合分析模型，假設不同臨床研究的差異不可忽略，baseline effect為固定效應，在整個網絡中，以處置間相對效應進行比較；Baseline and treatment effect model在Lu Ades model的基礎上，放寬baseline effect為隨機效應；而Arm-based model則假設不同臨床研究所比較的處置方式是被隨機分派到的，其結果的差異完全是由隨機誤差造成的，在整個網絡中，以處置的絕對效應進行比較。
因此本研究首先從理論上描述上述三種模型，針對Lu Ades model，斷裂的網絡需要額外獲取外部資料來重新連接，因此還需探討結合臨床試驗和觀察性研究的方法；隨後使用與實際資料評估不同模型的表現；針對模型在實際資料應用中發现的問題，進一步使用電腦模擬數據探討三種模型在網絡存在偏誤和異質性的情況下的表現；最後，結合模型理論特點和應用上的表現綜合評估分析，得出運用三種模型的注意事項及最適情形。

結論
Lu Ades model受限于fixed baseline effect，所以一旦統合分析中的網絡發生“斷裂”，則無法納入斷掉的資訊，需要人為尋找外部證據進行斷點連接，且Lu Ades model對用於連接的資料中存在的偏誤極為敏感，并會將偏誤傳遞到不含基準處理的network中，所以使用時應重點考量所納入的外部證據的品質。Baseline and treatment effect model和Arm-based model在baseline effect的假設上做了重大改變，將經典的fixed effect變為random effect，因此可以直接應用於斷裂的網絡上。一旦網絡中任何一個研究存在偏誤，Baseline and treatment effect model 便會講此偏誤傳播到整個network之中，雖然傳遞出去的偏誤很小，但是在應用中亦需非常警惕。Arm-based model更是將網絡中用於比較的對象由treatment contrast變為treatment response，假設納入的研究彼此之間無差異，相當於在整個網絡的層面上做了一次隨機分派臨床試驗，不容易在實際中找到能滿足其假設的情境。

關鍵字

網絡統合分析 ； 斷裂證據網絡 ； 相對治療效果模型 ； 基線模型 ； 絕對治療效果模型 ； 貝氏方法

並列摘要

Background
Compared to traditional pairwise meta-analysis, Network meta-analysis (NMA) can compare all treatments simultaneously and indirectly estimate two treatments that never be compared. This feature meets the need of modern evidence-based medicine, so NMA has a very rapid development recently.
For a convincing conclusion, researchers are always seeking to use all the information available. However, the evidence in NMA can be disconnected due to many reasons: it can be the single arm trial that results from toxicity testing or limit sample size of patients. Besides, when a “paradigm shift” happens, where new interventions are compared only to newer ones, abandoning the traditional treatment network. The disconnected network also arises when there are many standards of care due to the various medical condition in different areas, and when some particular trial endpoints differ from the commonly reported ones. Finally, in some studies, unpublished data can also contribute to the disconnected network.
To reconnect the network, researchers can contact to authors for complete raw data or wait for new data in on-going trials come into being. When these methods are unable to form a complete evidence base, one has to turn to specialized models.
Objectives
This thesis used three Bayesian modeling approaches to join the disconnected network, and evaluated their advantages and drawbacks theoretically and practically, then drew a conclusion on the best suitable conditions for each model.

Methods
Lu Ades model，Baseline and treatment effect model, Arm-based model were compared in this study. Lu Ades model is the classic approach for NMA. It assumes that relative treatment effects are exchangeable in the network, and a fixed baseline effect is applied to represent study effects. Baseline and treatment effect model uses relative treatment effects as its estimate scale as well, but it releases baseline assumption from fixed effect to random effect. Arm-based model makes a very strong assumption that treatment responses are exchangeable across the network, so the discrepancy of treatment response between studies originates from random sampling error.
This thesis first described three models in theory, then used osteoarthritis data to test model performance. Different simulation scenarios were also used to test models’ bias sensitivity and heterogeneity stability.

Results
Because of the fixed baseline effect, Lu Ades model does not have any flexibility to estimate disconnected network. Therefore, researchers have to find external sources such as well conducted observational studies to connect the network. Lu Ades model is also very susceptible to the bias in external evidence, and the bias in the pathway will transmit to the network which does not contain the reference treatment, so particular attention should be paid on bias introduced by non-randomized control trial. Baseline and treatment effect model assumes a random baseline effect, which gives it the flexibility to analysis disconnected network. When we apply Baseline and treatment effect model, any biased node in the network will affect the estimation of all the other treatments, although the magnitude of the bias is small. Arm-based model assumes that treatment responses are exchangeable in the network, and it is the least constrained model. When there are biased nodes in the network, the bias will only affect these problematic nodes, and the estimation of other treatments will remain unbiased. However, the assumption of Arm-based model is too strong, which is hard to be satisfied in reality.

並列關鍵字

Network meta-analysis ； Disconnected network ； Contrast-based model ； Baseline model ； Arm-based model ； Bayesian method

參考文獻

1. Zoccai, G.B., Network meta-analysis. 2014: Nova Science Publishers, Inc.

2. Falissard, B., et al., Real medical benefit assessed by indirect comparison. Therapie, 2009. 64(3): p. 225-32.

3. Efthimiou, O., et al., GetReal in network meta-analysis: a review of the methodology. Res Synth Methods, 2016. 7(3): p. 236-63.

4. Goring, S.M., et al., Disconnected by design: analytic approach in treatment networks having no common comparator. Res Synth Methods, 2016. 7(4): p. 420-432.

5. Administration, U.S.F.a.D. Designating an orphan product: drugs and biological products. 2016.5.2 [cited 2017 4.18]; Available from: https://www.fda.gov/forindustry/developingproductsforrarediseasesconditions/howtoapplyfororphanproductdesignation/default.htm.